Matrix Inverse

Oct 12, 2020
6 min read
Nov 24, 2020 05:10 UTC
...for a nonsingular matrix. We talk about left and right inverses, *the* matrix inverse and orthogonal matrices.
Mirrored alleyway in Switzerland.
Mirrored alleyway in Switzerland.

Linear system

We have matrix $\boldsymbol{A}: m \times n$, and vectors $\boldsymbol{b}: m \times 1$ and $\boldsymbol{x}: n \times 1$ where $\boldsymbol{b}$ is given and $\boldsymbol{x}$ is unknown. We use the equation

$$ \boldsymbol{Ax} = \boldsymbol{b} $$

We’re interested in when does a solution $\boldsymbol{x}$ exist, and how do we find it.

A linear system is consistent when there exists at least one $\boldsymbol{x}$ satisfying the equation. We can see that the following statements are equivalent:

  1. A linear system is consistent.
  2. $\boldsymbol{b} \in \mathcal{C}(\boldsymbol{A})$.
  3. $\mathcal{C}([\boldsymbol{A}, \boldsymbol{b}]) = \mathcal{C}(\boldsymbol{A})$.
  4. $r([\boldsymbol{A}, \boldsymbol{b}]) = r(\boldsymbol{A})$.

If $\boldsymbol{A}$ has full row rank $m$, then $\mathcal{C}(\boldsymbol{A}) = \mathbb{R}^m$, and $\boldsymbol{Ax} = \boldsymbol{b}$ is consistent for any $\boldsymbol{b}$.

The next fact is important in statistical contexts. The linear system $\boldsymbol{A}'\boldsymbol{Ax} = \boldsymbol{A}'\boldsymbol{b}$ is consistent for any $\boldsymbol{A}$, $\boldsymbol{b}$. This is because $\boldsymbol{A}'\boldsymbol{b} \in \mathcal{C}(\boldsymbol{A}') = \mathcal{C}(\boldsymbol{A}'\boldsymbol{A})$.

If $\boldsymbol{A}$ has full column rank, $\boldsymbol{A}'\boldsymbol{A}$ is nonsingular, and $\boldsymbol{x}$ is unique. This is an assumption in our regression analyses, which is there’s no redundancy in our predictor variables.

Matrix inverse

The inverse of a matrix only makes sense when the matrix is square. Before talking about it, we introduce two other concepts for non-square matrices.

Left and right inverses

If $\boldsymbol{A}: m \times n$, the right inverse $\boldsymbol{R}$ is the $n \times m$ matrix such that

$$ \boldsymbol{AR} = \boldsymbol{I}_m $$

The right inverse exists when $\boldsymbol{A}$ has full row rank, i.e. $r(\boldsymbol{A}) = m$ ($m \leq n$). This is because if $r(\boldsymbol{A}) = m$, then $\boldsymbol{Ax} = \boldsymbol{b}$ is consistent for any $\boldsymbol{b}$. Let

$$ \boldsymbol{b}_1 = \begin{pmatrix} 1 \\ 0 \\ 0 \\ \vdots \\ 0 \end{pmatrix}_{m \times 1}, \boldsymbol{b}_2 = \begin{pmatrix} 0 \\ 1 \\ 0 \\ \vdots \\ 0 \end{pmatrix}, \cdots, \boldsymbol{b}_m = \begin{pmatrix} 0 \\ 0 \\ 0 \\ \vdots \\ 1 \end{pmatrix} $$

The solution for $\boldsymbol{Ab}_1 = \boldsymbol{x}$ is $\boldsymbol{x}_1$, and similarly we have $\boldsymbol{x}_2, \cdots, \boldsymbol{x}_m$. If we put $\boldsymbol{x}_1, \cdots, \boldsymbol{x}_m$ into an $n \times m$ matrix

$$ \boldsymbol{R} = [\boldsymbol{x}_1, \cdots, \boldsymbol{x}_m], $$

we have $\boldsymbol{AR} = \boldsymbol{A}[\boldsymbol{x}_1, \cdots, \boldsymbol{x}_m] = \boldsymbol{I}_m$. From the proof we can also see that $\boldsymbol{R}$ is not unique.

Similarly, the left inverse $\boldsymbol{L}: n \times m$ is the matrix that satisfies $\boldsymbol{LA} = \boldsymbol{I}_n$. $\boldsymbol{L}$ exists when $\boldsymbol{A}$ has full column rank, i.e. $r(\boldsymbol{A}) = n (m \geq n)$.

Inverse

If $\boldsymbol{A}$ is non-singular, then its right inverse and left inverse exist, and they are the same, which we call the inverse of $\boldsymbol{A}$ denoted by $\boldsymbol{A}^{-1}$. $\boldsymbol{A}^{-1}$ is unique. To prove this, note that

$$ \boldsymbol{L} = \boldsymbol{LI}_n = \boldsymbol{LAR} = \boldsymbol{I}_n \boldsymbol{R} = \boldsymbol{R} $$

Properties

If $\boldsymbol{A}$ has full column rank $n$1, and there exists $\boldsymbol{L}$ such that $\boldsymbol{LA} = \boldsymbol{I}_n$ ($n \leq m$), then

  1. $\boldsymbol{B} = \boldsymbol{C} \Longleftrightarrow \boldsymbol{AB} = \boldsymbol{AC}$

    This is because

    $$ \boldsymbol{AB} = \boldsymbol{AC} \Rightarrow \boldsymbol{LAB} = \boldsymbol{LAC} \Rightarrow \boldsymbol{I}_n\boldsymbol{B} = \boldsymbol{I}_n \boldsymbol{C} $$

  2. For any $\boldsymbol{B}$, $r(\boldsymbol{AB}) = r(\boldsymbol{B})$. Pre-multiplying a full column rank matrix does not change the rank.

  3. For any $\boldsymbol{B}$, $\mathcal{R}(\boldsymbol{AB}) = \mathcal{R}(\boldsymbol{B})$ when $\boldsymbol{A}$ has full column rank.

If $\boldsymbol{A}$ has full column rank $n$, then $\boldsymbol{A}'\boldsymbol{A}$ is non-singular and vice versa. Similarly if $\boldsymbol{A}$ has full row rank, $\boldsymbol{AA}'$ is non-singular.

Inverse of non-singular matrix

Suppose $\boldsymbol{A}$ is an $m \times m$ non-singular matrix.

  • If $m = 2$, we have

    $$ \boldsymbol{A} = \begin{pmatrix} a & b \\
    c & d \end{pmatrix}, \quad \boldsymbol{A}^{-1} = \frac{1}{ad - bc} \begin{pmatrix} d & -b \\
    -c & a \end{pmatrix} $$

    Check that $\boldsymbol{AA}^{-1} = \boldsymbol{I}_2$ to make sure.

  • If $\boldsymbol{A}$ is diagonal and all diagonals are non-zero,

    $$ \boldsymbol{A}^{-1} = diag \left(d_1^{-1}, \cdots, d_m^{-1} \right) $$

  • $\boldsymbol{Ax} = \boldsymbol{b}$ has a unique solution $\boldsymbol{x} = \boldsymbol{A}^{-1}\boldsymbol{b}$.

  • $(k\boldsymbol{A})^{-1} = \frac{1}{k}\boldsymbol{A}^{-1}$.

  • $\boldsymbol{A}'$ is also non-singular, and its inverse is

    $$ (\boldsymbol{A}')^{-1} = (\boldsymbol{A}^{-1})' $$

    This is because if we take the transpose of both sides of $\boldsymbol{AA}^{-1} = \boldsymbol{I}_m$, we get $(\boldsymbol{A}^{-1})'\boldsymbol{A}' = \boldsymbol{I}_m$. Since $\boldsymbol{A}'$ is non-singular, $(\boldsymbol{A}^{-1})'$ is the inverse of $\boldsymbol{A}'$.

  • $\boldsymbol{A}^{-1}$ is also non-singular, i.e. $r(\boldsymbol{A}^{-1}) = m$. This is because

    $$ m = r(\boldsymbol{AA}^{-1}) \leq r(\boldsymbol{A}^{-1}) \leq m $$

  • If $\boldsymbol{B}$ is also an $m \times m$ non-singular matrix, then $\boldsymbol{AB}$ is also non-singular, and $(\boldsymbol{AB})^{-1} = \boldsymbol{B}^{-1}\boldsymbol{A}^{-1}$. To check this claim,

    $$ (\boldsymbol{AB})(\boldsymbol{B}^{-1}\boldsymbol{A}^{-1}) = \boldsymbol{AI}_m\boldsymbol{A}^{-1} = \boldsymbol{AA}^{-1} = \boldsymbol{I}_m $$

    This could be generalized to the product of $k$ non-singular matrices: $(\boldsymbol{A}_1\boldsymbol{A}_2 \cdots \boldsymbol{A}_k)^{-1} = \boldsymbol{A}_k^{-1} \cdots \boldsymbol{A}_1^{-1}$. This is similar to the transpose of the product of matrices.

  • If $\boldsymbol{A}: m \times n$ and $r(\boldsymbol{A}) = r$; $\boldsymbol{B}: m \times m$ is non-singular, then $r(\boldsymbol{BA}) = r$. Similarly if $\boldsymbol{C}: n \times n$ is non-singular, $r(\boldsymbol{AC}) = r$. In fact, $\boldsymbol{B}$ only needs to have full column rank, and $\boldsymbol{C}$ only needs to have full row rank.

Orthonormal matrices

An orthogonal matrix, or orthonormal matrix to be more precise, is a square matrix whose columns and rows are orthonormal. One way to express this is

$$ \boldsymbol{AA}' = \boldsymbol{A}'\boldsymbol{A} = \boldsymbol{I}_n $$

By definition, the columns have to be orthonormal because

$$ \begin{gathered} \boldsymbol{a}_i'\boldsymbol{a}_i = 1 = \|\boldsymbol{a}_i \|^2 \\
\boldsymbol{a}_i'\boldsymbol{a}_j = 0 \end{gathered} $$

Equivalently, $\boldsymbol{A}$ is orthogonal if $\boldsymbol{A}^{-1} = \boldsymbol{A}'$.

Examples

In general, any $2 \times 2$ orthogonal matrix $\boldsymbol{P}$ is

$$ \boldsymbol{P} = \begin{pmatrix} \sin\theta & -\cos\theta \\
\cos\theta & \sin\theta \end{pmatrix} $$

If we think of the matrix as an operator on a vector $\boldsymbol{u}$, the vector $\boldsymbol{v} = \boldsymbol{Pu}$ can be interpreted as a rotation of $\boldsymbol{u}$ by $\theta$ degrees. This is why $\boldsymbol{P}$ is sometimes called the rotation matrix. Note that this operation applies to higher-order matrices as well. For example, in PCA we can enforce zero-correlation by rotation.

Illustration of rotating a vector in $\mathbb{R}^2$ space.

The identity matrix is also obviously orthogonal. We can also permute the columns and the resulting matrix is still orthogonal. Geometrically, this is permuting the coordinate axes.

Properties

  • The product of two orthogonal matrices is orthogonal.

    To prove this, let $\boldsymbol{P}$ and $\boldsymbol{Q}$ be two orthogonal matrices. We then have

    $$ (\boldsymbol{PQ})(\boldsymbol{PQ})' = \boldsymbol{PQ}\boldsymbol{Q}'\boldsymbol{P}' = \boldsymbol{PP}' = \boldsymbol{I} $$

  • For any vector $\boldsymbol{x} \in \mathbb{R}^n$, $\| \boldsymbol{Px} \|^2 = \|\boldsymbol{x}\|^2$. Pre-multiplying an orthogonal matrix to a vector doesn’t change the norm of the vector because

    $$ (\boldsymbol{Px})'(\boldsymbol{Px}) = \boldsymbol{x}'\boldsymbol{P}'\boldsymbol{Px} = \boldsymbol{x}'\boldsymbol{x} = \|\boldsymbol{x}\|^2 $$

  • For any vector $\boldsymbol{x}, \boldsymbol{y} \in \mathbb{R}^n$, their inner product is not changed by pre-multiplying an orthogonal matrix

    $$ (\boldsymbol{Px})'(\boldsymbol{Py}) = \boldsymbol{x}'\boldsymbol{y} $$


  1. We can make similar arguments if $\boldsymbol{A}$ has full row rank. Just consider switching the orders. ↩︎


Related Posts